Applying computational modeling to drug discovery and development
Kumar et al. (2006): Applying computational modeling to drug discovery and development
This paper discusses "pharmaceutically-relevant computational modeling approaches currently used as predictive tools" and provides examples that "demonstrate how companies can employ these computational models to improve the efficiency of transforming targets into therapies".
The authors "have identified three areas where computational modeling has potential to substantially impact efficiency and development":
The first area is cell-signal behavior, where the application of models characterizes how lead compounds affect intracellular signaling. The second area is signal-response behavior, where models predict cellular phenotype from signaling information. The third area is physiology, in which models are used to simulate clinical outcomes. Each class of model can help identify new drug targets.
To demonstrate the interplay between traditional biology and high-throughout informatics, the authors provide the following example:
[T]he construction of a signaling model begins with an assembly of molecular interactions, rate parameters and spatial restrictions. Informatics groups analyze high-throughput datasets (ie. gene chip arrays, gene sequencing results, mass spectrometry results and yeast two-hybrid results) using methods like clustering or spacing alignments, and integrate results with data from other in-house biological experiments and from literature (obtained by text mining). The data are then further organized into ontologies. A model is constructed from a subset of these data and is then validated using traditional biology experiments. If the model captures experimental trends, it is used to generate predictions or hypotheses that suggest new biological experiments. The results of these experiments either further validate the model or identify novel biology that is then incorporated into the model. This interplay between informatics, modeling and traditional biology enables the focused use of large datasets to solve biologically relevant problems.
The next chapters deal with the three aforementioned areas - cell signaling models, signal-response models and physiological models.
The motivation for dealing with cell signaling models is explained by the authors with the following words:
Defects in signal transduction underlie many diseases that are of interest to pharmaceutical companies. For example, dysregulation of conserved protein tyrosine kinase pathways leads to a variety of cancers. Individual signaling proteins inside the cell are often the target of small-molecule drugs, whereas many antibody drugs target the receptors controlling signaling cascades.
How are such models constructed? The authors explain:
Typically, ordinary differential equations (ODEs) are used to describe mass-action kinetics and system behavior. Experimental measurement of reaction rates, concentrations, molecular interactions and trafficking parameters are essential for the construction of such models. The level of detail necessary varies from system to system, but many signal-transduction pathways can be modeled using a combination of measured values, fitted parameters, and coarse-grained descriptions of interactions.
According to the authors, "[m]odels that describe signaling pathways are important in pharmaceutical research for three main reasons":
(i) they often capture nonintuitive signal behavior and identify novel molecular function; (ii) they allow researchers to experiment in silico across a wide range of conditions (e.g. receptor numbers, ligand concentrations and phosphorylation rates), thus saving experimental resources and identifying important further experiments; and (iii) they serve as a database for much of the known information about a particular pathway.
As examples, the authors cite papers which describe the modeling of the Wnt pathway and the ErbB receptor.
The chapter on signal-response models starts with an interesting remark:
Interestingly, it has been hypothesized that no more than 20 signal transduction cascades control the seemingly endless list of cell behaviors observed in humans.
They further elaborate on this and come to the conclusion:
[T]o correct aberrant cellular behavior with drugs requires quantitative knowledge about multiple signaling proteins (that is, multivariate datasets). Multivariate datasets can then be used to understand cellular decision-making processes in the context of computational models. [...] Whereas ODE-level models are becoming more prevalent for describing signaling pathways, there are very few models that can accurately connect signaling pathways to cellular behavior at this level of mathematical description. The problem, therefore, requires the use of more abstracted signaling models. Abstracted models identify statistical relationships between signals and behavior, which suggest causal signal behavior relationships that can be further probed using molecular biology or genetic approaches.
As an example, a paper is cited which is about an investigation of "the molecular effects of an acute promyelocytic leukemia cell line treated with retinoic acid and arsenic trixoide", trying to answer the question how "downstream signaling events coordinate a known program of differentiation and apoptosis". Furthermore, the authors mention a "procedure based on linear modeling (partial least squares regression), whereby 8000 intracellular signals were correlated with more than 1000 apoptosis-related cellular responses".
In the final chapter on physiology, Noble's model of the human heart is mentioned.
The authors conclude:
Computational models address a key issue in the pharmaceutical industry: prediction. [...] The in silico component in research must still be coupled with hypothesis-driven experimental design and is not a substitute for the more important in cerebro component. [...] We believe that the most successful models will not only provide predictive power but will also be scalable, meaning that models currently appropriate for different phases in the R&D pipeline should be mutually compatible in anticipation of information that will connect disparate R&D stages.